The Library
Deep learning for high dynamic range imaging
Tools
Marnerides, Demetris (2019) Deep learning for high dynamic range imaging. PhD thesis, University of Warwick.
|
PDF
WRAP_Theses_Marnerides_2019.pdf - Submitted Version - Requires a PDF viewer. Download (119Mb) | Preview |
Official URL: http://webcat.warwick.ac.uk/record=b3452948~S15
Abstract
High Dynamic Range (HDR) imaging enables us to capture, manipulate and reproduce real world lighting with high fidelity. HDR displays are becoming more common, however, most content, from over 100 years of media, is Low Dynamic Range (LDR). Dynamic range expansion methods generate HDR from LDR content, for displaying on HDR displays and various other applications, attempting to recover missing information from the original HDR signal that is lost due to saturation or quantisation. Multiple methods have been proposed, addressing different aspects of the problem, however most are model-driven and require adjustment of parameters which may also vary depending on the content.
This thesis proposes using Convolutional Neural Networks (CNNs) for fully data-driven, end-to-end dynamic range expansion. A novel multibranch CNN termed ExpandNet is proposed that processes LDR inputs on multiple scales without using upsampling layers, which have been observed to cause artefacts when used in other traditional CNNs, in particular the UNet architecture. ExpandNet is evaluated and compared against traditional methods including other CNNs and is found to outperform all other methods on multiple metrics. An investigation of the effect of upsampling layers on the spectrum of the network outputs is then presented, characterising their impact in the Fourier domain, providing a way to assess the structural biases of CNNs. A new upsampling module is then proposed, based on the Guided Image Filter that provides spectrally consistent outputs when used in a UNet architecture, forming the Guided UNet (GUNet). The GUNet architecture is evaluated similarly to ExpandNet and is found to perform well, while executing faster and consuming less memory than ExpandNet. Finally, multiple methods using Generative Adversarial Networks (GAN) are proposed and evaluated, that hallucinate content in badly-exposed areas of the input while simultaneously expanding the range of the well-exposed areas. A single network configuration based on UNet, termed L-GAN, produces better results qualitatively and also performs well quantitatively compared to state-of-the-art methods such as ExpandNet and GUNet.
Item Type: | Thesis (PhD) | ||||
---|---|---|---|---|---|
Subjects: | Q Science > Q Science (General) Q Science > QA Mathematics > QA76 Electronic computers. Computer science. Computer software T Technology > TR Photography |
||||
Library of Congress Subject Headings (LCSH): | High dynamic range imaging, Neural networks (Computer science), Machine learning | ||||
Official Date: | June 2019 | ||||
Dates: |
|
||||
Institution: | University of Warwick | ||||
Theses Department: | School of Engineering | ||||
Thesis Type: | PhD | ||||
Publication Status: | Unpublished | ||||
Supervisor(s)/Advisor: | Debattista, Kurt, 1975- ; Khovanov, Igor | ||||
Sponsors: | Engineering and Physical Sciences Research Council | ||||
Format of File: | |||||
Extent: | xiv, 148, xxxv leaves : illustrations | ||||
Language: | eng |
Request changes or add full text files to a record
Repository staff actions (login required)
View Item |
Downloads
Downloads per month over past year